agent-context-kit
Context a coding agent can find on its own
agent-context-kit is three open Agent Skills that give a coding agent persistent, discoverable context about a codebase, so it works from what the project has written down instead of re-deriving your conventions, procedures, and automation from scratch every session. It is open source and free, and it is the same kit I use daily across my own projects.
I wrote up the thinking behind the kit in Best Practices for Coding Agents Start With a Legible Codebase.
The three skills
Each skill covers one kind of project knowledge an agent otherwise has to guess at, and each stores it in an obvious folder the agent is told to read before it starts work.
directives
The decisions your team has already made and does not want reopened: commits use imperative mood, every service runs in a container, build artifacts stay out of the source tree. They live as short markdown files in a project's docs/ folders, each with a one-line description so the agent can tell what's relevant without reading every file in full.
The agent loads the matching conventions before a task, so it follows a settled decision instead of guessing at it or re-litigating it in review. The skill also ships a small set of general-purpose defaults; a project's own docs take precedence over them.
how-tos
The operational procedures that otherwise live in one engineer's head: how to cut a release, how to generate an access token, how to run the deploy. Each is written once as an ordered, followable list in a how-tos/ folder.
The agent reads the procedure before it attempts the task, rather than reconstructing the steps from memory and getting one subtly wrong. Unlike directives, this skill ships no defaults, because operational procedures are specific to each project.
scripts
Discovery for the automation a project already has. It uses each script's own --help output as the source of truth for what the script does, not a comment header that drifts the moment someone changes the behaviour and forgets to update the comment.
With that in place the agent finds and runs what you already built, instead of quietly reimplementing it inline, a little differently, every session. It carries one requirement worth having anyway: every script must implement --help properly, since discovery runs it on each script it considers.
The three compose. An automatable how-to step points at a script rather than restating the command, and a directive about your git workflow governs how the tooling commits its own changes. Written together they stop being three separate piles of documentation and become one surface the agent reads the right slice of for whatever it is doing.
Why it exists
Most of what makes a coding agent effective on a real, non-trivial codebase is not the model. It is whether the codebase is legible to the agent: whether the conventions are written where it looks, whether the procedures are documented instead of remembered, whether the automation is discoverable instead of silently reimplemented. A stronger model shortens the time the agent spends thinking; it does nothing about the time it spends re-deriving facts about your project that were never written down.
These three skills are the minimum mechanism to make a codebase legible in that sense. They came out of a personal multi-project workspace where they get exercised every day, which is the only reason I trust them. The longer argument for why legibility beats prompt-tuning, worked through convention by convention, is in Best Practices for Coding Agents Start With a Legible Codebase. This page is the product; that post is the case for it.
Works with any agent, on any OS
SKILL.md is an open format, the Agent Skills specification,
not one vendor's proprietary layout. The same three skills work with
Claude Code, Codex, Cursor, OpenCode, and other agents that read it. They
run on plain Python 3.9+ and Git, identically on Linux, macOS, and
Windows, with no bash or WSL required, so the kit behaves the same
wherever your team works.
Install it
Install with npx skills add, which auto-detects the agents
you have installed, or through the Claude Code plugin marketplace the
repository self-hosts. Both routes, the exact commands and options, and
how to wire the AGENTS.md reminder block into your own project
live in the repository, which is the maintained source for all of it, so I
don't copy them here where they would fall out of date.